#!/bin/bash
# v14 = v10_r2 + generator-conditional prompt augmentation.
#
# Hypothesis: v11-v13b all regressed because reward-shape changes contaminated
# the single-seg majority (81% of data). v14 changes INPUT INFORMATION, not
# reward — for FORENSICS_GENCOND_PROB of training samples, prepend "The forged
# segments were generated by <gen>." to the prompt. GRPO then organically
# strengthens generator-specific low-level fingerprint pathways because the
# reward (iou+strict_boundary+strict_edge, unchanged from v10_r2) is higher
# when the model exploits those features. The remaining (1-p) generic samples
# force the unconditional path to also benefit from those learned features.
#
# Target: wan (47.74) and vace-1.3B (51.58) — currently 18-22pt below ltx
# (70.30). These are the hardest distributions and have the most room.
#
# Eval-time default is FORENSICS_GENCOND_MODE=none (matches deployment). Run
# the 3-way eval script to also measure the correct-name oracle and the
# wrong-name negative control.
#
# Only difference vs run_grpo_forensics_v10_r2.sh: FORENSICS_GENCOND_PROB,
# OUTDIR, LOG_PATH, WANDB_NAME, master_port.

export WANDB_PROJECT=Forensics-GRPO
export WANDB_NAME="${WANDB_NAME:-AF_v14_gencond}"

export PYTHONPATH=".:$PYTHONPATH"
export DEBUG_MODE="true"
export LOG_PATH="./qwen2.5_7b_vl_forensics_v14_gencond.txt"

export PATH="/mnt/local-fast/zhangt/torch_env/bin:$PATH"
export LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"

# === v10_r2 baseline (unchanged) ===========================================
export FORENSICS_COT=false
export FORENSICS_FORGERY_AWARE=false

export FORENSICS_HUNGARIAN_DECOMP=true
export FORENSICS_HUNGARIAN_MAXK=8
export FORENSICS_BOUNDARY_WEIGHT=0.3
export FORENSICS_BOUNDARY_TAU=2.0

export FORENSICS_SPI_AUG=true
export FORENSICS_SPI_PROB=0.25
export FORENSICS_SPI_CHUNK_S=2.5
export FORENSICS_SPI_SAFETY_S=0.5

export FORENSICS_FBR_AUG=true
export FORENSICS_FBR_PROB=0.15

export FORENSICS_STRICT_TAUS="0.3,0.5,0.7"
export FORENSICS_STRICT_EDGE_SIGMA=0.5

unset FORENSICS_COT_CONSIS_WEIGHT
unset FORENSICS_COT_CF_WEIGHT
unset FORENSICS_COT_ALIGN_WEIGHT
unset FORENSICS_K_REQUIRE_TIMESTEP
unset FORENSICS_IOU_REQUIRE_THINK
# ============================================================================

# === v14: generator-conditional prompt =====================================
# 50/50: half the steps see the generator name (conditioning), half don't
# (transfer / generic-path training). At eval time we run generic-only and
# also probe correct-name / wrong-name to diagnose how the signal propagates.
export FORENSICS_GENCOND_PROB="${FORENSICS_GENCOND_PROB:-0.5}"
# ============================================================================

OUTDIR="${OUTDIR:-outputs_forensics/v14_gencond}"
MODEL_PATH="${MODEL_PATH:-/mnt/local-fast/zhangt/Qwen2.5-VL-7B-Instruct}"
PREPROCESSED="${PREPROCESSED:-/mnt/local-fast/zhangt/forensics_grpo_cache_uniform3584_fps2.0}"

cd "$(dirname "$0")/.."

torchrun --nproc_per_node="8" \
    --nnodes="1" \
    --node_rank="0" \
    --master_addr="127.0.0.1" \
    --master_port="12374" \
    src/open_r1/grpo_forensics.py \
    --deepspeed scripts/zero3_offload.json \
    --output_dir "$OUTDIR" \
    --model_name_or_path "$MODEL_PATH" \
    --annot_dir /mnt/local-fast/zhangt/annot/annot \
    --video_root /mnt/local-fast/zhangt/video \
    --preprocessed_data_path "$PREPROCESSED" \
    --reward_funcs iou strict_boundary strict_edge \
    --dataset_name activityforensics \
    --max_prompt_length 8192 \
    --max_completion_length 512 \
    --num_generations 4 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --logging_steps 1 \
    --bf16 \
    --torch_dtype bfloat16 \
    --data_seed 42 \
    --gradient_checkpointing true \
    --attn_implementation flash_attention_2 \
    --num_train_epochs 4 \
    --run_name "$WANDB_NAME" \
    --report_to none \
    --save_steps 200 \
    --save_total_limit 4 \
    --learning_rate 1e-6 \
    --save_only_model true